Staged Training Report ✓ Complete

Run ID: shoulder_session_multiheight_decoder_only_20_no_mask_ratio
Generated: 2026-02-25 15:32:38
Stages Completed: 1
Total Elapsed Time: 04:33:53

Configuration

No config defaults changed since last commit.

All Staged Training Parameters (61 parameters)

Parameter	Value
`total_samples`	10000000
`batch_size`	8
`stage_samples_multiplier`	100000000000
`update_interval`	250
`window_size`	100
`num_best_models_to_keep`	1
`sampling_mode`	Loss-weighted
`loss_weight_temperature`	0.5
`loss_weight_refresh_interval`	50
`stop_on_divergence`	True
`divergence_gap`	0.002
`divergence_ratio`	1.5
`divergence_patience`	50
`divergence_min_updates`	10
`val_spike_threshold`	2.0
`val_spike_window`	15
`val_spike_frequency`	0.75
`val_plateau_patience`	250
`val_plateau_min_delta`	0.0001
`custom_lr`	0.0001
`disable_lr_scaling`	True
`custom_warmup`	-1
`lr_min_ratio`	0.001
`resume_warmup_ratio`	0.05
`plateau_factor`	0.8
`plateau_patience`	15
`preserve_optimizer`	False
`preserve_scheduler`	True
`samples_mode`	Train additional samples
`num_random_obs_to_visualize`	2
`selected_frame_offset`	3
`runs_per_stage`	5
`serial_runs`	True
`clean_old_checkpoints`	True
`enable_baseline`	False
`baseline_runs_per_stage`	1
`run_id`	shoulder_session_multiheight_decoder_only_20_no_mask_ratio
`seed`	None
`enable_wandb`	True
`wandb_project`	developmental-robot-movement
`lr_sweep.lr_min`	1e-07
`lr_sweep.lr_max`	0.01
`lr_sweep.phase_a_num_candidates`	5
`lr_sweep.phase_a_seeds`	1
`lr_sweep.phase_a_time_budget_min`	3.0
`lr_sweep.phase_a_survivor_count`	2
`lr_sweep.phase_b_seeds`	3
`lr_sweep.phase_b_time_budget_min`	10.0
`lr_sweep.ranking_metric`	median_best_val
`lr_sweep.min_samples_before_timeout`	1000
`lr_sweep.min_evals_before_stop`	5
`lr_sweep.save_sweep_state`	True
`plateau_sweep.enabled`	True
`plateau_sweep.plateau_ema_alpha`	0.85
`plateau_sweep.plateau_improvement_threshold`	0.0015
`plateau_sweep.plateau_patience`	25
`plateau_sweep.cooldown_updates`	5
`plateau_sweep.max_sweeps_per_stage`	2
`plateau_sweep.min_sweep_improvement`	0.0
`initial_sweep_enabled`	True
`stage_time_budget_min`	180

World Model Architecture (config.py)

Parameter	Value
`AUTOENCODER_LR`	0.0002
`BATCH_SIZE`	1
`CANVAS_HISTORY_SIZE`	3
`DECODER_DEPTH`	8
`DECODER_EMBED_DIM`	256
`DECODER_NUM_HEADS`	8
`DECODER_ONLY_DEPTH`	20
`EMBED_DIM`	256
`ENCODER_DEPTH`	5
`FOCAL_BETA`	5
`FOCAL_LOSS_ALPHA`	0.1
`FRAME_SIZE`	(224, 224)
`GRADIO_UPDATE_INTERVAL`	1
`LR_MIN_RATIO`	0.001
`MODEL_TYPE`	decoder_only
`NUM_HEADS`	4
`PATCH_SIZE`	16
`PERCEPTUAL_LOSS_WEIGHT`	0
`SEPARATOR_WIDTH`	16
`WARMUP_STEPS`	1000
`WEIGHT_DECAY`	0.01
`MASK_RATIO_MIN`	1
`MASK_RATIO_MAX`	1
`TRAIN_MASK_RATIO_MIN`	1.0
`TRAIN_MASK_RATIO_MAX`	1.0

Timing Summary

Stage	Plateau Sweeps	Sweep Time	Training Time	Stage Total
Stage 1	2	00:35:56	00:11:43	00:47:39
TOTAL	2	00:35:56	00:11:43	00:47:39

Initial LR Sweep: Stage 1: selected LR 3.16e-05 in 00:17:13

Plateau Sweep Details

Total Sweeps: 2
Stages with Sweeps: 1 of 1
Total Sweep Time: 00:35:56
Average Sweep Duration: 00:17:58

Stage 1: 2 sweeps

LR Progression: 3.2e-05 → 3.2e-05 → 1.8e-06

Sweep #	Triggered At (samples)	Wall Time	Selected LR	Duration
1	14,592	00:04:13	3.16e-05	00:19:26
2	26,624	00:27:07	1.78e-06	00:16:30

Stage Results

Stage	Best Loss	Stop Reason	Samples Trained	Time	Sweeps	LR (Initial→Final)
Stage 1	0.116476	max_sweeps (2)	13,824	00:47:39	2	3.2e-05→1.8e-06

Total Plateau Sweeps: 2

Stop Reason Breakdown

max_sweeps: 1 stages

Loss Across Full Training Run

Multi-Run Statistics

Total Runs: 5
Average Best Loss: 0.119031 ± 0.002157
Best Overall: 0.116476
Worst Overall: 0.122095

Stage 1 (5 runs)

Run	Best Loss	Stop Reason	Samples	Time	Selected
1	0.120900	max_sweeps (2)	11,008	01:01:41
2	0.117126	max_sweeps (2)	11,520	00:49:31
3	0.118558	max_sweeps (2)	13,568	00:48:28
4	0.122095	max_sweeps (2)	9,472	00:46:07
5	0.116476	max_sweeps (2)	13,824	00:47:39	✓
Mean: 0.119031 ± 0.002157		Min: 0.116476 / Max: 0.122095		Range: 0.005619

Best Checkpoint

Name: best_model_auto_session_so101_multiheight_part1_1345_shoulder_session_multiheight_decoder_only_20_no_mask_ratio_00034048_cont_val_0.116476.pth
Stage: 1
Hybrid Loss (full session): 0.130061

Learning Rate Timeline with Plateau Sweeps

Stage Progression

Stage	Orig Loss	Train Loss	Time	Samples	Stop Reason
1 ⭐	0.130061	0.116476	00:47:39	13824	max_sweeps (2)

Hybrid Loss Over Original Session (per Stage)

Stage 1 (Best) - Hybrid Loss: 0.130061

Sample Counts

Cumulative Across All Stages

Per Stage

Stage 1 (Best) - Total Samples: 13,824

Best Checkpoint Inference

Selected Frame 3

Action 0

Action 1

Action 2

Random Observations

Observation 550

Action 0

Action 1

Action 2

Observation 918

Action 0

Action 1

Action 2

Individual Stage Reports

Stage 1 Run 5 (Staged) (orig loss: 0.130061) ⭐